update

app.py

@@ -113,17 +113,17 @@ def prepare_pipeline(model_version, enable_realism, enable_anti_blur):
             pipeline.infusenet_sim.cpu()
             pipeline.image_proj_model_sim.cpu()
             torch.cuda.empty_cache()
-            pipeline.infusenet_aes.to(pipeline.pipe.device)
+            pipeline.infusenet_aes.to('cuda')
             pipeline.pipe.controlnet = pipeline.infusenet_aes
-            pipeline.image_proj_model_aes.to(pipeline.pipe.device)
+            pipeline.image_proj_model_aes.to('cuda')
             pipeline.image_proj_model = pipeline.image_proj_model_aes
         else:
             pipeline.infusenet_aes.cpu()
             pipeline.image_proj_model_aes.cpu()
             torch.cuda.empty_cache()
-            pipeline.infusenet_sim.to(pipeline.pipe.device)
+            pipeline.infusenet_sim.to('cuda')
             pipeline.pipe.controlnet = pipeline.infusenet_sim
-            pipeline.image_proj_model_sim.to(pipeline.pipe.device)
+            pipeline.image_proj_model_sim.to('cuda')
             pipeline.image_proj_model = pipeline.image_proj_model_sim
 
         loaded_pipeline_config['pipeline'] = pipeline

pipelines/pipeline_infu_flux.py

@@ -309,9 +309,10 @@ class InfUFluxPipeline:
         face_info = sorted(face_info, key=lambda x:(x['bbox'][2]-x['bbox'][0])*(x['bbox'][3]-x['bbox'][1]))[-1] # only use the maximum face
         landmark = face_info['kps']
         id_embed = extract_arcface_bgr_embedding(id_image_cv2, landmark, self.arcface_model)
-        id_embed = id_embed.clone().unsqueeze(0).float().cuda()
+        id_embed = id_embed.clone().unsqueeze(0).float()
         id_embed = id_embed.reshape([1, -1, 512])
         id_embed = id_embed.to(device='cuda', dtype=torch.bfloat16)
+        self.image_proj_model.to('cuda', torch.bfloat16)
         with torch.no_grad():
             id_embed = self.image_proj_model(id_embed)
             bs_embed, seq_len, _ = id_embed.shape